LLM-friendly knowledge representation

Package source knowledge into task-specific artifacts instead of handing the model raw enterprise sprawl: Dropbox normalizes files to markdown and builds graph-derived knowledge bundles; Uber converts Google Doc tables to markdown and stores document titles/summaries/FAQs; Pinterest compiles table schemas, table summaries, and historical queries; DoorDash builds per-domain review profiles; LinkedIn encodes organizational procedures as playbooks; Wix uses AI-optimized docs/skills and feedback-generated knowledge or prompt-instruction documents; Culligan Quench organizes enablement content around sales use.

Persist represented knowledge in retrieval stores, usually embeddings/vector stores, and often pair that with lexical or metadata retrieval: Dropbox stores embeddings, chunks, and contextual graph representations and uses BM25 plus dense vectors; Pinterest generates embeddings for table summaries and historical queries in OpenSearch; Uber indexes chunks into a vector store and also adds BM25; LinkedIn stores conversation embeddings alongside chronological memory; Wix adds embeddings for new knowledge documents and retrieves top-scoring documents from a vector database.

Narrow the model’s context to the subset relevant to the task, domain, source, or user intent: DoorDash routes PRs to only relevant review rules; Uber QueryGPT maps prompts to business-domain workspaces and prunes schemas; Uber Genie uses source identification to restrict retrieval; Pinterest retrieves top-N tables and asks an LLM to choose top-K; LinkedIn hides thousands of tools behind a small set of meta-tools to reduce context bloat; Wix tests curated skills or MCP-packaged skills as constrained context.

Use LLMs to create or enrich the knowledge representation before retrieval or generation: Uber converts extracted table contents into markdown tables; Pinterest forwards prompts to an LLM to create table/query summaries and uses an LLM to select relevant tables; Wix uses LLMs for feature extraction from owner feedback; LinkedIn uses LLMs to summarize learned patterns, extract episodic activities, and compress conversational memory.

Make represented knowledge updateable rather than static: Grab auto-updates users’ Knowledge Vaults; Wix classifies owner feedback into new knowledge, prompt instructions, or nothing and stores new documents plus embeddings; LinkedIn asynchronously indexes user activity and periodically compresses conversational memory; Dropbox stages graph relationships asynchronously before creating knowledge bundles.

Expose knowledge representations through tools, workflows, or agent interfaces when the task requires action: Dropbox wraps its index as a “super tool”; LinkedIn exposes playbooks as MCP tools; Grab combines LLM calls, Python execution, and Knowledge Vault lookups in structured workflows; DoorDash runs review/fix agents in remote VMs with full repository context.